Generalizable Blind Image Quality Assessment with Retrieval Augmentation and Contrastive Vision-Text Local Meta-Training
Year: 2026
Type: conference-paper
Abstract: While existing CLIP-based Image Quality Assessment (IQA) models have demonstrated enhanced capabilities in perceiving visual quality, due to the small scale of IQA datasets, they still struggle to evaluate out-of-distribution distorted images, which deviate from the pre-training datasets. Although some mixed-dataset training methods have been proposed to enable IQA models to be trained on more diverse datasets, the limited scale of IQA data still restricts these methods' generalization ability. To address these issues, we propose a novel CLIP-based blind IQA method named R etrieval- e nhanced loca l meta t raining (Relocat)-IQA, which aims to augment both of the image and text embeddings by online retrieval for better generalization. Specifically, we first propose an effective pre-training framework to align and encode the fine-grained multi-modal quality knowledge as an external embedding memory. Then the sample-specific quality knowledge is retrieved from the memory to enhance the image and text embeddings during the training of \method, which can greatly improve the zero-shot predictions at inference time. In addition, we propose a local meta-training strategy to utilize mixed datasets efficiently. Remarkably, we demonstrate the benefits of the retrieval-enhanced module for generalization not only through experiments but also from a theoretical perspective. Extensive experiments validate the effectiveness and generalization of our method. Supp. Mat. and details can be found at https://github.com/zhyan567/Relocat-IQA.
Institutions: Peking University, Chinese University of Hong Kong, Harbin Institute of Technology, University of Science and Technology of China, Chinese Academy of Sciences, Hefei Institutes of Physical Science, The University of Texas at Arlington, City University of Hong Kong, State Key Laboratory of Media Convergence and Communication, Communication University of China
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225319665